False memories for scenes using the DRM paradigm.

Area of Science:

• Cognitive Psychology
• Computer Vision
• Neuroscience

Background:

• Human memory for photographs is highly accurate.
• Understanding stored visual representations is key to memory research.
• Quantifying visual similarity in stimuli is crucial for experimental design.

Purpose of the Study:

• To explore convolutional neural networks (CNNs) as a measure of visual similarity for complex scenes.
• To assess if CNN-derived similarity predicts human perceptual and memory performance.
• To validate CNNs as a tool for future visual memory studies.

Main Methods:

• Experiment 1: Participants identified the most visually dissimilar image within a category, using CNN-defined image space.
• Experiment 2: A modified Deese-Roediger-McDermott paradigm tested memory for photographs preselected by their distance to a visual prototype (CNN centroid).
• Utilized CNNs to define image space and measure perceptual/representational similarity.

Main Results:

• CNN-based visual similarity correlated with human behavioral outcomes.
• Participants could detect visually distant "odd-one-out" scenes.
• False alarm rates in recognition tests increased for scenes closer to the visual prototype.

Conclusions:

• CNNs effectively measure visual scene similarity relevant to human perception and memory.
• The findings support the use of CNNs for quantifying stimuli in visual memory research.
• This approach offers a novel method for studying memory for complex visual information.